问题
I am trying to build a neural network which would multiply 2 numbers. To do the same, I took help of scikit-learn. I am going for a neural network with 2 hidden layers, (5, 3) and ReLU as my activation function.
I have defined my MLPRegressor as follows:
X = data.drop('Product', axis=1)
y = data['Product']
X_train, X_test, y_train, y_test = train_test_split(X, y)
scaler = StandardScaler()
scaler.fit(X_train)
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test)
mlp = MLPRegressor(hidden_layer_sizes=(5, 3), activation="relu", learning_rate="adaptive", max_iter=500000, verbose=True, validation_fraction=0.25)
Here, data is the dataframe which contains the 3 columns, 2 random numbers and 1 Product column.
The issue is the loss that I get is of order 10^14. How do I reduce this loss, improve my model performance and what all possible changes can help me out in this situation?
回答1:
I believe that an NN with only two hidden layers is not sufficient to perform a multiplication of arbitrary numbers. Note that a multiplication of N * M is equal to the M-fold addition of N. There are neural networks that can perform an addition, but you have to consider the general conditions. Furthermore, you have to decide how the inputs and outputs of your network are defined: Do you want two input neurons and one output neuron, or do you want the two multiplicants to be binary (as a vector) in the net? I think you would have to normalize your input values if you gave decimal values in two input neurons.
回答2:
I am not an expert in NNs. I would do log-tranform of the inputs, then feed them into a network, then exponetiate the output. Just a thought.
回答3:
There is a way to do this. This is a neural network that approximate the multiplication function to multiply numbers in the range of 1000 - 10000, and it could work really well if the range of numbers is limited. Here is the gist link
来源：https://stackoverflow.com/questions/52660944/how-to-build-a-neural-network-to-multiply-two-numbers